Distinguishing Among 3-D Distributions for Brain Image Data Classification

To facilitate the process of discovering brain structure-function associations from image and clinical data, we have developed classification tools for brain image data that are based on measures of dissimilarity between probability distributions. We propose statistical as well as non-statistical methods for classifying three dimensional probability distributions of regions of interest (ROIs) in brain images. The statistical methods are based on computing the Mahalanobis distance and Kullback-Leibler distance between a new subject and historic data sets related to each considered class. The new subject is predicted to belong to the class corresponding to the dataset that has the smaller distance from the given subject. The non-statistical methods consist of a sequence of partitioning the brain image into hyper-rectangles followed by applying supervised neural network models. Experiments performed on synthetic data representing mixtures of nine distributions as well as on realistic brain lesion distributions from a study of attentiondeficit hyperactivity disorder (ADHD) after closed head injury showed that all proposed methods are capable of providing accurate classification of the subjects with the Kullback-Leibler distance being the least sensitive on the size of the training set and on information about the new subject. The proposed statistical methods provide comparable classification to neural networks with appropriately generated attributes, while requiring less computational time.